Tesla has just held a streaming event called AI Day, and alongside the new humanoid robot that the company plans to launch soon, there has been a lot of talk about Fully Autonomous Driving. For now, the device is not technically a true autonomous driving system – but if we are to believe Elon Musk, the device will evolve. With the promise that he will one day be able to drive safer than human drivers.
“I am confident that our hardware 3 Full Self Driving computer 1 will be able to deliver fully autonomous driving with much better levels of safety than when driving is provided by a human. The system will be at least 200% or 300% better. Then, of course, there will be the hardware 4 FSD computer 2 which we will probably be launching with the Cybertruck – so probably within a year or so. He will be about four times more capable ”, explains Elon Musk.
Tesla believes fully autonomous driving will be much safer than human drivers
Tesla then went into detail. For example, we saw how the car “sees” the road, other users and pedestrians in what the company calls Vector Space. It is in fact a simple 3D representation compiled from eight cameras via artificial intelligence. In the demo that you can see at the end of the article, the Vector Space is displayed on the Tesla console.
This representation knows how to overcome problems such as the temporary obstruction of the video stream. The AI can indeed predict the position and direction of everything around it even if the sensors temporarily no longer see certain objects and people likely to cross the trajectory of the Tesla. It works as well at intersections as on the highway for example.
The data from the Vector Space is then sent to what Tesla calls the Neural Net Planner. It is this component that actually manages the driving. There is a collection of algorithms doped with AI. As a bonus, it is responsible for simulating the consequences of each driving decision several thousand times per minute – taking into account the trajectory predictions of other users and pedestrians.
The system works so well that the device now knows how to make fairly natural decisions. In one example, the FSD system can be seen changing its decision in the face of an indecisive rider – from a situation of giving way to resuming driving when the computer has determined that the rider in question will let it go.
Tesla has developed a new chip in-house to improve the training of its AI
All of this is calculated in real time in the vehicle. Yet the car doesn’t get smarter the more you drive it. The real secret to how Fully Autonomous Driving works is the AI models developed by Tesla – which are already ready to use. These require a lot of work, because it has been necessary to train them with vast amounts of reliable data, and to test them relentlessly, all on supercomputers.
An effort that continues to be made by Tesla’s R&D teams on an ongoing basis. Ultimately, increasing the reliability of the system will also mean increasing the computing power. For this, Tesla unveiled Project Dojo, an initiative to internally develop a new chip adapted to the ambitions of the firm.
The first chip in question, Tesla D1 specializes in distributed computing. D1 chips are packaged in numbers in what Tesla calls ExaPOD, a computing unit that can deliver up to 1.1 exaflops. Which is the equivalent of the combined power of approximately 30,500 Nvidia RTX 3090 graphics cards.
For now, there is no question of integrating ExaPODs directly into Tesla. The system is cumbersome to say the least: it occupies about ten mainframe computer cabinets. But this increased computing power will directly benefit the precision of machine learning models and therefore the reliability of Fully Autonomous Driving.
You can watch the full presentation, in English, in the video below: